Emma Kerr, Deimos Space UK Ltd.; Elisabeth Geistere Petersen, Deimos Space UK Ltd.; Patrick Talon, Deimos Space UK Ltd.; David Petit, Deimos Space UK Ltd.; Chris Dorn, Inverse Quanta Ltd; Stuart Eves, SJE Space Ltd
Keywords: characterisation, light curve, AI, machine learning, GEO, simulation
Abstract:
The classification and characterisation of objects in geosynchronous orbit (GEO) is an important goal in Space Situational Awareness (SSA). Optical observations can be used to detect and identify GEO objects, while radar is generally only used to observe Low Earth Orbit due to the range limitations. Temporal variations of apparent magnitude, called a light curve, are captured by optical telescopes. This optical signature is usually processed with various estimation algorithms to get insightful information on the state of the object. Light curves contain information on features such as attitude, size, shape and materials useful for the characterisation of the object. Analysing light curves can allow the determination of normal patterns of life, and hence the detection of anomalous behaviour during the orbit such as manoeuvres.
The main objective of this study is to develop a machine learning algorithm for extrapolating common features of GEO objects. In order to achieve this goal a high-fidelity simulator was developed to generate photometric data and light curves to be further processed and used to train the machine learning algorithm. The simulated observations are intended to reproduce images taken by Deimos Sky Survey (DeSS) system, located in Puertollano (Spain). This paper details the development and intermediate results of the machine learning algorithm.
Using simulated and real light curve data, an artificial neural network has been trained to identify characteristics of spacecraft in GEO. The algorithm uses a combination of unsupervised, as well as supervised, CNN model architectures to learn and extract patterns from the target object light curves. First, an autoencoder was trained to extract meaningful information from the light curves. The encoder part of the model was then used as feature extractor in classification models. Depending on the quality of the observed light curve, the classification algorithm can extract a range of properties: object shape, size, materials and attitude. In previous work, a method using four different models trained to identify these individual characteristics was outlined. However, to improve efficiency a multi-branch classifier with an independent output for each characteristic has been designed to replace the individually trained models. This paper will discuss the design, implementation and results of the multi-branch model and next steps for the machine learning development.
The light curve simulator developed for the project can be used to generate a full night light curve for an orbiting object. By varying the simulator inputs, many different nights, under different conditions and for different objects can be simulated. A wide dataset helps the algorithm to recognise a broad range of satellite types, attitudes and many other characteristics. Development and validation of the algorithm is ongoing, using simulated light curves and real data gathered by DeSS. This work is intended to be the first step in automatic characterisation of all space objects in all orbits.
Date of Conference: September 14-17, 2021
Track: Non-Resolved Object Characterization